Citadel Securities believes the AI market is showing signs of increasing demand elasticity. Since the end of June, usage-weighted token prices have fallen by about 40%, but H100 rental prices have rebounded from their June lows; in July, the top 1% of companies spending AI saw an average monthly increase of about 49% in spending per employee. Hyperscale cloud vendors' revenue reached $106.3 billion in the second quarter, a year-on-year increase of about 43%, indicating that new computing power can still quickly find paying demand.
The standards for measuring the economics of AI are also changing. The cost of completing an effective task is more relevant than the price of a single token.
Sufficiently capable open-source models allow customers to reduce costs while maintaining near-perfect output quality, but this increases AI usage, supports GPU and cloud service demand, and weakens the pricing power of closed-source cutting-edge models.
Citadel also cautions that the computing power pricing market is still young and opaque; current high utilization rates do not guarantee supply and demand balance in the coming years. Once model architectures significantly improve efficiency, computing power demand may fall short of expectations implied by current capital expenditures.
